Your Insurance Company's AI Is Denying Claims With a 90% Error Rate. Here's What Real Computer Use AI Actually Looks Like.

The $25.7 Billion Problem Nobody Wants to Talk About

A February 2025 survey from Premier Inc. put a number on something every hospital CFO already knew in their gut. Claims adjudication costs healthcare providers $25.7 billion annually. Of that, $18 billion is potentially unnecessary, meaning it exists purely because of bad processes, manual rework, and systems that don't talk to each other. That's not inefficiency. That's a choice. Every time a claims adjuster manually re-enters data from a PDF into a portal, every time someone calls to check a claim status that an automated system could surface in seconds, every time a denial gets appealed because the original review was done by a fatigued human on their 40th claim of the day, that's a billable hour someone is paying for. And it's a billable hour that a proper computer use agent could eliminate entirely. The industry has known about this waste for years. The reason it persists isn't technical. It's organizational inertia, vendor lock-in, and frankly a quiet comfort with the status quo among the people who benefit from billing those hours.

UnitedHealth's AI Disaster Is a Warning, Not a Template

Let's be direct about what happened with UnitedHealth and their nH Predict algorithm. Multiple lawsuits, confirmed by federal court proceedings in early 2025, alleged that UnitedHealthcare used an AI model with a 90% error rate to systematically deny Medicare Advantage claims for elderly patients. The kicker, and this is the part that should make your blood pressure spike, is that the lawsuit alleged they knew about the error rate and kept using it anyway. Why? Because they also knew that most patients wouldn't appeal. The system wasn't broken. It was working exactly as designed, just not for patients. This is what happens when 'AI automation' means 'automate the denials and let humans deal with the fallout.' KFF data from January 2026 shows Medicare Advantage insurers made nearly 53 million prior authorization determinations in 2024. Fifty-three million. If even a fraction of those were influenced by flawed models, we're talking about millions of people whose care was delayed or denied by software that nobody outside the insurer ever audited. The lesson here isn't that AI shouldn't touch insurance claims. The lesson is that AI being used to obscure decisions from humans is a fundamentally different thing from AI being used to help humans make better, faster decisions.

Why Old-School RPA Is Dying and What's Killing It

• ●Traditional RPA tools like UiPath and Automation Anywhere work by following rigid, pre-programmed scripts. Change one field label in a web portal and the whole bot breaks. Insurance portals change constantly.
• ●Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by end of 2027, largely because teams are deploying glorified chatbots and calling them agents. That's not a computer use agent. That's a FAQ page with delusions of grandeur.
• ●RPA can't handle unstructured data. A scanned EOB, a handwritten doctor's note, a PDF with a non-standard layout. These are the exact documents that clog claims pipelines, and they're exactly what rule-based bots choke on.
• ●The average insurance claim touches 7 to 10 different systems before it's resolved. RPA requires a separate bot script for each system. A real computer use agent navigates all of them the same way a human does, by seeing the screen and deciding what to do next.
• ●Maintenance costs for RPA deployments in insurance can run 30 to 50% of initial build costs annually. You're not automating your way out of the problem. You're adding a new layer of fragility on top of the old one.

What 'Real' Computer Use AI Actually Does in a Claims Workflow

Here's the thing that separates a genuine computer use agent from the RPA tools and the denial-machine algorithms that are getting insurers sued. A real computer use agent sees your desktop the way a human does. It reads screens, navigates portals, opens PDFs, fills forms, cross-references data across systems, and makes judgment calls based on what it actually sees. Not API calls. Not pre-mapped fields. Actual visual computer use. Think about what that means for a claims workflow. An AI computer use agent can log into a provider portal, pull the claim, cross-reference it against policy documents, check for prior authorization status, flag discrepancies, and update the internal system, all without a human touching it. And when something genuinely needs human judgment, it stops, surfaces the relevant context, and hands off cleanly. That's not science fiction. That's what the best computer use agents are doing right now. The benchmark that matters here is OSWorld, the industry standard test for real-world computer task completion. Most AI models are still struggling to break 60% on it. Claude Sonnet 4.5 from Anthropic, one of the stronger models, scores 61.4%. The gap between 60% and production-ready is enormous in a claims environment where errors cost money and sometimes cost people their healthcare.

Why Coasty Is the Answer This Industry Actually Needs

I'm not going to pretend there's a long list of viable options here, because there isn't. Coasty.ai scores 82% on OSWorld. That's not a marketing number. That's the benchmark score, and it's higher than every competitor right now. In a claims processing context, that gap matters enormously. A 60% OSWorld score means the agent fails on 4 out of 10 real-world computer tasks. In insurance, those failures are rework, errors, compliance risk, and angry policyholders. An 82% score means the agent handles the overwhelming majority of real desktop workflows without hand-holding. Coasty controls actual desktops, browsers, and terminals. It's not an API wrapper pretending to be an agent. It can navigate the exact portals your adjusters use every day, the ones that don't have APIs, the ones that require a human to click through six screens to get a status update. The desktop app works with your existing infrastructure. Cloud VMs let you run parallel workloads. Agent swarms mean you can process claims in parallel instead of sequentially, which is where the real speed gains are. And there's a free tier, so you don't need to convince a procurement committee before you can see whether it actually works. BYOK support means your sensitive claims data doesn't have to live in someone else's infrastructure if your compliance team has opinions about that, and they always do.